fix: Cohere catalog omits active models returned by GET /v1/models

[github-actions]

Summary

Adds missing Cohere catalog entries that can be represented under the current schema, based on first-party docs and model cards.

Cohere’s live catalog still includes embed/rerank IDs that models.dev cannot express cleanly (vector/rerank outputs; see #572). This change covers the independently verifiable chat, ASR, and parse models only.

Changes

Lab metadata (models/cohere/)

ID	Notes
tiny-aya-global	3.35B multilingual chat; 8k/8k; text→text; open weights CC-BY-NC-4.0
tiny-aya-earth	Same limits; West Asia / Africa specialization
tiny-aya-fire	Same limits; South Asia specialization
tiny-aya-water	Same limits; Europe / Asia-Pacific specialization
cohere-transcribe-03-2026	2B ASR; audio→text; open weights Apache-2.0; limit 0/0 (peer ASR pattern)
parse-v5.0	Document image parse; image→text; context 8192; closed weights

Provider overrides (providers/cohere/models/)

ID	Provider fields
tiny-aya-* (4)	base_model only; cost omitted (no separate public per-token list)
cohere-transcribe-03-2026	base_model only; cost omitted (trial free / Model Vault hourly)
parse-v5.0	base_model only; cost omitted (page-billed, not $/MTok)

Intentionally not added

• All embed-* / *-image IDs
• All rerank-* IDs
• Reason: modalities enum is text|audio|image|video|pdf only; no vector/rerank output (issue + Seperate Embedding Category #572)

Evidence

Claim	Source	Establishes
Tiny Aya IDs, Chat endpoint, 8k context/output, 70 languages, regional specializations	https://docs.cohere.com/docs/models · https://docs.cohere.com/docs/tiny-aya	Live chat models and limits
Tiny Aya launch 2026-02-17, open weights	https://cohere.com/blog/cohere-labs-tiny-aya · https://huggingface.co/CohereLabs/tiny-aya-global	Release date, license CC-BY-NC-4.0, HF weights
Transcribe ID, ASR, 14 languages, Apache 2.0, free trial API / Model Vault prod	https://docs.cohere.com/v2/docs/transcribe · https://cohere.com/blog/transcribe · https://huggingface.co/CohereLabs/cohere-transcribe-03-2026	Model facts, pricing shape, release 2026-03-26
Transcribe API: audio file in, text out, optional temperature	https://docs.cohere.com/v2/reference/create-audio-transcription	Modalities + temperature
Parse ID Live, context 8192, image→structured text/markdown	https://docs.cohere.com/docs/models · https://docs.cohere.com/docs/parse · https://docs.cohere.com/v2/reference/parse	Model + API shape; API currently image_url only
Embed/rerank still documented but schema-blocked	https://docs.cohere.com/docs/models · https://docs.cohere.com/docs/rerank · #572	Why those IDs were skipped

Validation

• Cross-checked fields against the first-party URLs above (no live GET /v1/models call; no bun validate run in this environment).
• Matched patterns to existing Cohere lab/provider files and peer ASR entries (e.g. Whisper / Voxtral: limit 0/0, optional cost).

Review notes

• Parse release_date / last_updated: set to 2026-08-27 (verification day). Cohere marks parse-v5.0 Live but does not publish a launch date in the docs checked. Prefer replacing with a first-party launch date when available.
• Tiny Aya / Transcribe / Parse costs: omitted rather than guessing. Aya Expanse FAQ pricing ($0.50/$1.50) was not applied to Tiny Aya. Transcribe production is Model Vault hourly; Parse is page-billed.
• Embed/rerank: left out pending modalities support (Seperate Embedding Category #572); not a TOML-only fix.
• cohere-transcribe-arabic-07-2026: documented as Live on the models page but not in the issue missing list; not added here.

Closes #5628

Automated by the issue fixer: https://github.com/anomalyco/models.dev/actions/runs/33050679728